The Autonomous Agent Loop

Every pattern so far had a shape drawn by your code. The autonomous agent replaces the drawn shape with a single loop: plan → act via a tool → observe the result → verify against the goal → repeat. The loop is the whole program — "sophisticated behavior, simple implementation."

The animation runs a coding task around this cycle. Each lap, the agent plans a step, fires a tool call — terminal and file icons animate real actions — then reads what the environment actually returned. That reading is the crucial edge: ground truth from the environment is what separates an agent from a model compounding its own guesses. Watch the error lap: a command fails, and instead of repeating it, the agent reads the failure output and self-corrects on the next pass.

Watch the step counter climb — nobody predicted the lap count, which is exactly why this task couldn't be a fixed workflow. And watch the exit: a stopping condition (goal verified, or max iterations reached) breaks the loop with the completed task. Stopping conditions, checkpoints, and sandboxes are what make the autonomy trustworthy — the trade for all this flexibility is cost, latency, and the risk of compounding errors, so agents earn their keep only on open-ended problems where no path could be hardcoded.